Conformal Prediction for Causal Effects of Continuous Treatments

Maresa Schröder, Dennis Frauen, Jonas Schweisthal, Konstantin Heß, Valentyn Melnychuk, Stefan Feuerriegel

TL;DR

This paper introduces the first conformal prediction framework for estimating the causal effects of continuous treatments (like drug dosages) with finite-sample guarantees. Unlike previous methods that only handle binary "yes/no" treatments or require perfectly known treatment policies, this approach provides reliable uncertainty intervals even when the treatment policy must be estimated from noisy, real-world data.

What Problem Does This Paper Address?

In safety-critical fields like personalized medicine, knowing the average effect of a treatment isn't enough. A doctor needs to know the range of possible outcomes for a specific patient to assess risk. While Machine Learning (ML) is great at providing point estimates, it often struggles with Uncertainty Quantification (UQ).

Standard UQ methods often fail in causal inference because of a "distribution shift." When we intervene (e.g., change a patient's dosage), we break the natural relationship between patient characteristics and the treatment they received—a relationship captured by the propensity score. This violation of "exchangeability" makes standard prediction intervals invalid.

Existing solutions for this problem are mostly limited to binary treatments (e.g., "Drug A" vs. "Placebo"). Applying them to continuous treatments (e.g., "10mg" vs "15mg") is non-trivial because you rarely observe two patients with the exact same continuous treatment value, and the math behind the distribution shift becomes significantly more complex.

Figure 1: CP intervals on finite-sample data. UQ methods with asymptotic guarantees might suffer from under-coverage and are often not faithful. Thus, we aim at CP with finite-sample guarantees.

Methods

The authors build on Split Conformal Prediction, a model-agnostic technique that uses a "calibration set" to determine how much to "stretch" a model's prediction to guarantee a certain level of coverage (e.g., 95%).

The paper tackles two main scenarios:

  1. Known Propensity Score: Used for "soft interventions" where we know the original treatment policy and want to see the effect of a slight shift (e.g., increasing everyone's dosage by 5mg).
  2. Unknown Propensity Score: The more common real-world case. Here, the authors must estimate the propensity score from data, which introduces additional uncertainty.

To handle continuous treatments, they employ kernel smoothing. Since we can't find exact matches for a specific dosage $a^*$, they use a Gaussian kernel to weight nearby observations. They also introduce a sensitivity parameter $M$ to bound the potential error in the estimated propensity score, ensuring the intervals remain "faithful" (valid) even if the propensity model isn't perfect.

The paper distinguishes between known policies (Scenario 1) and unknown policies where the intervention cannot be expressed as a simple function of the original (Scenario 2).

Datasets

The method was evaluated on three types of data:

  • Synthetic Datasets: Two environments with different propensity and outcome functions (one step-wise/concave, one Gaussian/oscillating) to test the math in a controlled setting.
  • Semi-synthetic (TCGA): Based on The Cancer Genome Atlas, using real gene expression data (4,000 features) but with simulated treatments and outcomes to allow for ground-truth benchmarking.
  • Real-world (MIMIC-III): A large database of ICU records. The authors looked at the duration of mechanical ventilation (continuous treatment) and its effect on patient blood pressure.

Key Insights and Ideas

  • Finite-Sample over Asymptotic: Most UQ methods only work "in the limit" (with infinite data). This method provides guarantees that hold for any sample size, which is vital for rare diseases or small clinical trials.
  • The "Unknown" is the Innovation: This is the first work to provide finite-sample guarantees for continuous treatments when the propensity score is unknown. As shown in Figure 2, they fill a significant gap in the literature.
  • Robustness to Shift: By reformulating the calibration as an optimization problem (specifically, a "Type-I invex" problem), they can mathematically "undo" the distribution shift caused by the intervention.
Figure 2: Key works on causal CP.

Key Results

The authors compared their method against common baselines like MC Dropout and Deep Ensembles.

  • Faithfulness: On synthetic data, the proposed method consistently hit the target coverage (e.g., 90% or 95%), while baselines often fell significantly short (sometimes as low as 10-20% coverage), meaning their "uncertainty" was dangerously overconfident.
  • Informativeness: On the TCGA dataset, their intervals were not just valid but also much narrower (and thus more useful) than those produced by simply binarizing the treatment and using older methods.
  • Real-world Utility: In the MIMIC-III medical data, the intervals correctly showed higher uncertainty in treatment regions that were rare in the training data (e.g., very long ventilation durations), a critical feature for clinical decision support.

Limitations and Future Work

  • Propensity Quality: While the method accounts for estimation error, very poor propensity models will lead to excessively wide (conservative) intervals that may not be useful for decision-making.
  • Computational Cost: The method requires solving an optimization problem for each prediction. While the authors found it manageable (averaging ~16 seconds on a desktop CPU), it might be slow for large-scale real-time applications.
  • Univariate Outcomes: Currently, the framework is designed for a single continuous treatment and a single outcome. Extending this to multiple treatments or multi-dimensional outcomes remains an open challenge.

Summary updated

Abstract

Uncertainty quantification of causal effects is crucial for safety-critical applications such as personalized medicine. A powerful approach for this is conformal prediction, which has several practical benefits due to model-agnostic finite-sample guarantees. Yet, existing methods for conformal prediction of causal effects are limited to binary/discrete treatments and make highly restrictive assumptions such as known propensity scores. In this work, we provide a novel conformal prediction method for potential outcomes of continuous treatments. We account for the additional uncertainty introduced through propensity estimation so that our conformal prediction intervals are valid even if the propensity score is unknown. Our contributions are three-fold: (1) We derive finite-sample prediction intervals for potential outcomes of continuous treatments. (2) We provide an algorithm for calculating the derived intervals. (3) We demonstrate the effectiveness of the conformal prediction intervals in experiments on synthetic and real-world datasets. To the best of our knowledge, we are the first to propose conformal prediction for continuous treatments when the propensity score is unknown and must be estimated from data.
BibTeX
@inproceedings{schroder2025conformal,
  author = {Maresa Schröder and Dennis Frauen and Jonas Schweisthal and Konstantin Heß and Valentyn Melnychuk and Stefan Feuerriegel},
  title = {{Conformal Prediction for Causal Effects of Continuous Treatments}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=1nL84tQNnK},
}
Figure